Performance of string tokenisation: String.split() and StringTokenizer compared

As noted in our introduction to the String.split() method, the latter is around twice as slow as a bog-standard StringTokenizer, although it is more flexible.

Compiling the pattern

Using String.split() is convenient as you can tokenise and get the result in a single line. But it is sub-optimal in that it must recompile the regular expression each time. A possible gain is to compile the pattern once, then call Pattern.split():

Pattern p = Pattern.compile("\\s+");
...
String[] toks = p.split(str);

However, the benefit of compiling the pattern becomes marginal the more tokens there are in the string being tokenised. The graph below shows the time taken on one test system (2GHz Pentium running JDK 1.6.0 under Windows) to tokenise 10,000 random strings consisting of between 1 and 19 5-character tokens separated by a single space¹.

Now, what these results highlight is that whatever method you use, there's also no need to panic unless you're doing a lot of tokenisation: even the fairly extreme case of tokenising 10,000 19-token strings took just 60 milliseconds. However, if you are doing a serious amount of tokenising (for example, in some of my corpus linguistics work, I frequently process corpora consisting of millions of words and occasionally shaving time off the "nuts and bolts" of parsing can be useful), then the good old StringTokenizer may still be worth using for performance reasons. If you can cope with its restrictions, of course. The results above suggest that compiling the pattern is comparatively most beneficial if you are tokenising a large number of strings each with few tokens.

1. Each figure is an average of ten runs. The three tests were interleaved in each run, and 3 dummy runs preceded the timed runs to allow the JIT compiler to "warm up". The code to parse a single string was in each case coded as a separate method, so that this method would become "hot".